Jev - TypeSafe AI

Three Typed Question Primitives

TypeSafe exposes Choice, Score, and Noul, three primitive question types. Choice selects one option from a defined list, Score rates content against ordered levels, and Noul evaluates a yes or no statement. Each returns a typed answer with probabilities and confidence.

Parallel Evaluation

All questions in a single request are evaluated in parallel against the same state. Adding more questions barely changes response time, and independent evaluation avoids context-rot between questions.

Calibrated Confidence and Probabilities

Every Choice and Score answer includes a confidence estimate and a probability distribution. Software can branch, sort, and route on these values, and escalate to human review when confidence is low.

Zero-Hallucination, Type-Safe Outputs

Possible outputs are defined in advance, and the text-generation layer is skipped entirely. The model never makes type errors and cannot produce content outside the declared structure, which TypeSafe describes as zero hallucination.

Machine-Native Speed and Cost

Jev evaluates every question against the state in a single query using a parallel sampler. TypeSafe positions it as roughly two orders of magnitude faster and cheaper than existing LLMs on System One tasks, at 42 USD per billion input tokens.

• Support teams: Classify incoming tickets by department, urgency, and frustration with three parallel questions in one call.

• Content platforms: Screen messages entering and leaving an LLM app, thresholding hazard probabilities to pass, review, block, or route.

• Customer service bots: Classify user intent, then route each request to deterministic logic, a specialist LLM, or a human handler.

• Data teams: Extract structured fields such as dates, emails, and amounts from documents, resolving candidates in code with confidence-based review.

• Search engineers: Score retrieved passages or candidate pairs for relevance and rank results before they reach an answering model.

What are System One Models and what is Jev?

System One Models are a new class of AI model built for decisions inside software rather than for chat. Jev is TypeSafe's first public System One Model, optimized for automation. You send Jev structured questions together with a piece of state, and it returns typed decisions with probabilities and confidence that your code can act on directly.

Is Jev just a smaller LLM?

No. Jev uses a different architecture, a parallel sampler, and a different training method called Reinforcement Learning for Calibrated Decisions (RLCD). It gives up string generation entirely and is optimized for structured, type-safe outputs, which is why it is much faster and cheaper than LLMs on decision tasks.

How is this different from JSON mode or structured outputs?

JSON mode still generates free-form text that you must parse and validate, and the model can still produce malformed or hallucinated content. Jev skips the text-generation layer, so possible outputs and their structure are defined in advance denying the model room to hallucinate or produce type errors.

How can Jev be so fast and inexpensive?

Jev evaluates every question against the state in a single parallel query instead of generating tokens sequentially. This parallel sampler plus a lean architecture make it roughly two orders of magnitude faster and cheaper, benchmarked at 193.6x faster and 444.6x cheaper than LLM workflows on System One tasks.

Can Jev still get things wrong?

Yes. Jev returns calibrated probabilities and confidence rather than a guarantee of correctness, so answers carry honest uncertainty estimates. When confidence is low, your software is expected to escalate to a reviewer or a different model, which is the intended design of the system.

How do I get started?

Create a free account at console.typesafe.ai, generate an API key, and open the Playground to test questions against sample state. For code, send a POST request to the systemone endpoint with a model alias such as jev-latest, or install the Python or JavaScript SDK.

• Open the Playground at console.typesafe.ai and sign in to test Jev with sample text and mixed question types without writing code.

• Create an account and generate an API key from the dashboard. TypeSafe is open to everyone, so no waitlist or approval is required.

• Send a POST request to the systemone endpoint with state, a model name such as jev-latest, and typed questions describing the decisions your code needs.

• Read the structured response directly in code. Each answer includes the chosen option, score, or noul value with probabilities and confidence, ready to branch on.

• Install a client SDK to simplify integration. The Python package is typesafe-sdk and the JavaScript SDK is available from the TypeSafe registry, with typed request and response objects.

• Add the TypeSafe agent skill to Claude Code or Codex to let an agent use System One questions while building, using a single install command.